github

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.85). The skill uses gh to fetch GitHub issue/PR comments (including line-level review comment bodies) via REST/GraphQL at runtime; those comment bodies are authored by other users (outsiders) and are ingested as readable text into the agent’s LLM context.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (high risk: 0.90). The prompt explicitly tells the agent to "perform gh commands outside of the sandbox," which directs it to bypass sandbox/security boundaries and execute CLI operations on the host (potentially modifying system state), so it encourages compromising the machine's environment.